Generalize sensitive / triplet loss to multiple attributes

[e-pet]

Do a rather general implementation of the sensitive loss that works for multiple overlapping attributes with respect to which a model should be "invariant", e.g. chest drain presence / pacemaker presence / patient sex.